MIT CADR on-line help and documentation recovery

What “on-line help” means in this audit

On-line has its historical meaning here: documentation deliberately reachable from the running program. The audit includes a source construct when at least one of these conditions holds:

1. a runtime command explicitly opens or displays the file;
2. ZWEI registers it as command, mode, option, key, or self-documentation metadata;
3. the Lisp DOCUMENTATION path can retrieve it as a definition docstring or documentation property;
4. a window, menu, mouse, error handler, editor, or application implements an explicit Help, ?, :DOCUMENT, or :DOCUMENTATION endpoint;
5. it is a reviewed generated self-documentation artifact in the public source tree.

Comments, bug mail, change logs, and manuals with no identified in-program consumer are excluded. Generated comparison files are retained but labelled separately from a proven runtime display path.

How CADR help works

Recovered System 46 corpus

The extractor first selects the greatest numeric version for each directory-local, case-folded source stem. In the pinned tree this yields 463 source files. It parses historical Lisp bytes without evaluating forms, ignores comments and quoted data, and retains the surrounding top-level form for every matched declaration. One backquoted DEFCOM generator is retained as a template rather than misreported as a concrete installed command.

The parser requires a source docstring to be followed by an executable body form; a sole string body is a return value, not documentation. Help-character detection is also a reviewed pathname-and-definition allowlist. CADR sources use #/? and #\HELP as ordinary character data in music, punctuation, keyboard translation, cursor glyphs, and device acknowledgements, so a raw token search would create false Help handlers.

The 949 recovered records break down as follows:

Of those records, 825 contain literal documentation strings, 30 preserve nonliteral forms that compute or select documentation, and 94 are handlers whose bodies produce help at runtime. Categories overlap: for example, a menu's mouse handler can be both mouse documentation and an explicit Help endpoint.

The tracked asset directory contains 96 files totaling 2,003,292 bytes:

• LICENSE.source, copied from the pinned public source tree;
• catalog.json and the readable catalog.md inventory;
• 89 .help.lisp files containing only exact matched top-level source contexts, not silently reformatted whole source files;
• four byte-exact standalone or generated ZWEI artifacts.

Every catalog entry records original pathname, versioned filename, line and byte spans, source kind, and SHA-256. Tests verify that each catalogued context hashes to the bytes in its tracked recovery file.

Standalone and generated ZWEI artifacts

The two comparison files are documentation artifacts, but they are not promoted to the same evidence grade as basic.zwei's explicit consumer. Contemporary source-tree notes describe their contents; that does not by itself prove a command displayed them.

Reproducing the tracked recovery

From a checkout of the public source repository at commit 8e978d7d1704096a63edd4386a3b8326a2e584af:

python3 scripts/extract-cadr-help.py \
  --source /path/to/mit-cadr-system-software/src \
  --output docs/assets/mit-cadr-online-help \
  --clean

The extractor refuses a source tree whose copied license or four reviewed standalone artifacts do not match the pinned byte sizes and hashes. It additionally verifies all 463 selected source files: 12,024,129 bytes, with manifest SHA-256 efe818fc76d02c6c31afd6b0bbcdf43d89d277e871dedbee6562445be21d662d. The manifest hashes each file as sorted UTF-8 relative pathname, a NUL byte, and the binary SHA-256 file digest. The complete catalog and aggregate declaration counts are validated before --clean may replace a previous recovery.

The public source license is retained at LICENSE.source. The original is src/LICENSE.

LM-3 System 303 cross-check

The maintained LM-3 Fossil tree corroborates the source-integrated architecture and adds later material. At public system-303 check-in 4df393c68d7f083ce42d5c377039d26043cc18a9031ace28258dc97f4137eb91, the metadata-only scan finds 10,090 matching metadata records from 10,073 unique source forms in 562 files, plus eight standalone candidates. One source form can yield more than one record when it carries multiple recognized kinds. The scan covers later variable, structure, type, and SETF (DOCUMENTATION ...) metadata, DEFINE-COMMAND-DOCUMENTATION, and a reviewed 65-definition Help-key dispatch set in addition to the System 46 mechanisms. The inventory marks 188 generated-template records and labels patch contexts rather than claiming that every lexical declaration was active in one running world.

The later source makes the breadth of on-line documentation especially visible:

• 1,196 executable :DOCUMENTATION fields have a following non-NIL value. Three distribution/dist.lisp occurrences are boolean directory-classification flags, leaving 1,193 help fields: 405 structure-slot descriptions, 244 typed flavor descriptions, five Site Data Editor class descriptions, 532 menu or UI properties, and seven live documentation-message clauses.
• The positional producer audit finds 141 signal descriptions, four resource descriptions, 74 demo-menu descriptions, 29 Common Lisp-style CERROR continuation descriptions, 13 MULTIPLE-CERROR proceed choices, 14 System-key registrations, and 43 typeout-item descriptions.
• Forty-five DEFMETHOD operations contain a segmented DOCUMENTATION name. Twenty-one return a sole literal string and 24 compute their result. A typeout-window DEFWRAPPER, two :CASE-DOCUMENTATION clauses, the central who-line renderer, and a ZWEI command-documentation handler add computed endpoints. Three DEFSELECT who-line clauses include one literal and two computed results. The scan records these endpoints without treating helper or routing infrastructure as additional payload.
• TFRAME contributes 29 command mouse descriptions and 29 corresponding long :DOCUMENTATION fields, plus 16 option descriptions. ZMail contributes 94 top-level command descriptions, 29 who-line updaters, 13 static who-line strings, 17 option associations, and 95 user-option presentation records across generated and direct constructors.
• Application-specific UI channels add one inspector :BUTTON-DOCUMENTATION property and its nonliteral documentation-string builder; two ZMail who-line override init properties, two setter payloads, and two dynamic bindings; 12 grabbed-mouse who-line bindings; five FED special-command mouse strings; and four ZWEI global mouse-blinker documentation bindings. NIL clearing and saved-value restoration forms are excluded because they introduce no source documentation.
• Four source initializers have direct consumers: Converse option help, ZWEI search help, the ZMail filter summary string, and ZMail's command-documentation alist. The Converse initializer is a second payload in a DEFVAR whose ordinary variable docstring is already a separate record, so it increases the record count without adding another unique source form.
• Source occurrences are intentionally not deduplicated across applied-patch and main source paths. The 29 string-first CERROR forms, for example, include a two-form Site Data Editor pair present in both places; the inventory preserves that provenance.

These application-specific records are consumer-backed. The inspector reads BUTTON-DOCUMENTATION and turns it into its who-line result. ZMail's override mixin returns its dynamically bound or set string. The central renderer is both installed on the who-line field and shown selecting grabbed-mouse or window-method documentation in its runtime body. ZWEI's query-replace handler implements :NAME, :SHORT, and :FULL and is installed as a DOCUMENTATION-FUNCTION.

The segmented-name sweep deliberately does not count every function containing the word DOCUMENTATION. Flavor normalization only moves already inventoried flavor text into generic Lisp documentation storage; the wind/emack request builds an offline Bolio manual; and ZMail updater and append helpers orchestrate the already typed updater, association, and payload records. Likewise, grouped :CASE-DOCUMENTATION routing in method-combination machinery is not a second endpoint or source payload.

Across all mechanisms, 9,609 records point to literal strings, 190 to nonliteral source forms, and 291 to computed handlers. The pinned producer audit also records meaningful zeros: there are no DEFINE-MAIL-TEMPLATE or DEFSIGNAL-FORMAT invocations. All 33 DEFINE-SETF-METHOD occurrences—32 concrete forms and one generated template—omit a documentation string, and the three DEFINE-NOT-SETTABLE-MAIL-FILE-OPTION forms install bare option symbols rather than display or help metadata.

The parser does not invoke a Lisp reader. It nevertheless preserves source reader-object boundaries for nested block comments, reader conditionals, read-time forms, pathnames, vectors, structures, complex numbers, dimensioned arrays, radix-prefixed lists, and ZL package-prefixed forms. Files explicitly declaring Readtable:CL use backslash quoting; ZL and older System 46 sources retain slash quoting. This distinction is required both for TFRAME's quoted strings and for ZL format strings that legitimately end in a backslash.

The optional scan is pinned independently: its 1,044 eligible source and documentation files total 25,493,051 bytes and have manifest SHA-256 0758ba2f96f1764fe622d283e2f51277b5bf3060481b4019973767db5ad046ec under the same pathname-plus-file-digest convention as the System 46 manifest.

The standalone candidates are:

These are not all original MIT System 46 artifacts: the Fossil check-in is a maintained System 303 restoration, and its tape material includes a later LMI backport. The tree also contains conflicting redistribution signals: the restoration note describes a liberal license, while the included historical copyright file requires prior written LMI permission for further distribution. There is no repository-level license that clearly resolves every payload. For that reason the optional LM-3 pass writes only paths, hashes, counts, structural or symbolic identities, and non-text declaration metadata to ignored build/help/mit-cadr/lm3-inventory.json; it does not copy source, documentation text, menu labels, or demo titles. Literal names are replaced with kind-and-byte-offset identities before serialization.

Run that cross-check by adding:

python3 scripts/extract-cadr-help.py \
  --source /path/to/mit-cadr-system-software/src \
  --lm3-source /path/to/lm-3-system-303-checkout \
  --clean

Completeness boundary

The tracked count is complete for declarations matching the reviewed mechanisms in the pinned System 46 source snapshot. It is not a claim to reproduce a running load band exactly. Macro expansion, reader conditionals, runtime table mutation, site-local systems, and compiled-only FEF documentation can change what a particular machine displays. A future load-band audit should enumerate live command tables, keyboard repositories, properties, flavors, and compiled debugging records, then compare that state with this source recovery.

Sources

• MIT CADR System 46 source, commit 8e978d7d1704096a63edd4386a3b8326a2e584af, verified 2026-07-19.
• LM-3 sys Fossil repository, maintained System 303 check-in 4df393c68d7f083ce42d5c377039d26043cc18a9031ace28258dc97f4137eb91, verified 2026-07-19.